Multi-category rotated SAR ship generation with multi-input Generative Adversarial Networks
Moran Ju, Buniu Niu, Qing Xi Hu · Remote Sensing Letters · 2023
Deep learning-based Synthetic Aperture Radar (SAR) ship detection is a research hotspot at present due to its wide application in both military and civil fields. However, the lack of the labelled SAR ship data makes it a challenging task. To alleviate this problem, an improved multi-input Generative Adversarial Networks (MGAN) is proposed to generate the rotated SAR ship targets with its category and location labels. The proposed MGAN mainly consists of multi-input preprocessing, improved generator and self-supervised discriminator. To restrict the location, category and pixel value, the proposed MGAN takes the location, category and the pixel value constraints as input. In this way, plentiful SAR ship images with its rotated location and category labels can be obtained. Furthermore, we design category encoder and self-supervised constructor to improve the quality of the SAR ships generated by MGAN. The Intersection over Union (IoU) between the generated and the manual location labels reaches 0.83, which demonstrates the accuracy of the generated location label. The results of the experiment on classical oriented target detectors indicate that the SAR ships generated by MGAN can significantly improve the detection performance of oriented SAR ship targets.